Fun piece that only shows how easy it is to jailbreak AIs (meaning getting around "forbidden" queries, etc). The research was developed by a group that includes Anthropic as well as some universities.